Contrasting Identifying Assumptions of Average Causal Effects: Robustness and Semiparametric Efficiency
Tetiana Gorbach, Xavier de Luna, Juha Karvanen, Ingeborg Waernbaum; 24(197):1−65, 2023.
Abstract
This paper examines different identifying assumptions for estimating average causal effects using observational data. The three assumptions studied are the back-door assumption, which utilizes pre-treatment covariates, the front-door assumption, which involves mediators, and the two-door assumption, which combines pre-treatment covariates and mediators. The paper provides efficient influence functions and semiparametric efficiency bounds for each assumption and their combinations. It is shown that none of the identification models consistently provides the most efficient estimation, and conditions are given under which some bounds are lower than others. The paper also investigates the robustness of the estimators to misspecification of the nuisance models and presents simulation experiments to evaluate the finite sample behavior of the estimators. The findings have practical implications for analysts who need to choose between different identifying assumptions and corresponding estimators, as they highlight the trade-off between efficiency and robustness to model misspecification.
[abs]